qihongl / iterativelasso_mcw

run iterative lasso on the MCW data set

MCW_iterativeLasso

Run iterative lasso on the MCW data set. The iterative lasso code was adapted from knowledge and concepts lab github, which can be retrieved from this link: https://github.com/UWKCL/iterativelasso

intro the MCW dataset:

1200 rows

• 1-900: TrueWords

• 901 - 1200: non-words

• 1-94: TrueAnimals

• 95-240: TrueArtifacts

rowfilter and colfilter are filtered data. In particular, outlying voxels and voxels that never active (always 0) were removed.

Goal:

• classifcation: animals vs. artifacts, animals vs. non-animal words, animals vs. everything else, artifacts vs. non-artifacts words, etc.

notes 6/1:

data for subject1:

• X = numItems x numVoxels, matrix
• Y = numItems x 1, logical vector (e.g. 1 = 'animal', 0 = 'non-animal')

obejctive: Yhat = XB close to Y

• Since we are using LASSO, many elements in B are going to be zero out

lambda: sparsity parameter.

• bigger lambda ~> more zeros
• note that the "best lambda" is unknown

outer k folds cross validation

• maximize the "performance" on the test set

inner k folders cross validation

• tune lambda with respect to the "performance"

steps: e.g.

• for every subjects:
• do 10 folds CV
• with in each fold, try lambda1, lambda2, ..., lambdaN
• AND for each lambda, do 9 folds CV with cvglmnet()